你开心就好
Jetson-Orin-Nano-Super (3) 测试 JetPack
编辑
2025-07-27
记录知识
0
请注意，本文编写于 46 天前，最后修改于 33 天前，其中某些信息可能已经过时。

目录

确定支持范围
本地测试
测试deviceQuery
测试conv_sample
测试tensorrt
测试vpi
使用容器
安装镜像
测试deviceQuery
测试conv_sample
测试tensorrt
测试vpi
参考文献
总结

刚装好的jetson nano，需要简单的验证一下cuda nn，本文基于刚刷好的机器，进行cudann的简单测试

确定支持范围

nvidia jetpack 默认情况下cuda有一定的支持范围，我们需要先针对机器安装情况和官网的范围进行核对，如果不是在范围之内，则需要升级一下软件。否则出现如下错误。

 CUDA Device Query (Runtime API) version (CUDART static linking)

cudaGetDeviceCount returned 35
-> CUDA driver version is insufficient for CUDA runtime version
Result = FAIL

这类问题，首先需要查看nvidia-smi的输出，获取cuda版本和driver版本，如下

# nvidia-smi 
Sun Jul 27 14:54:34 2025       
+---------------------------------------------------------------------------------------+
| NVIDIA-SMI 540.4.0                Driver Version: 540.4.0      CUDA Version: 12.6     |
|-----------------------------------------+----------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |         Memory-Usage | GPU-Util  Compute M. |
|                                         |                      |               MIG M. |
|=========================================+======================+======================|
|   0  Orin (nvgpu)                  N/A  | N/A              N/A |                  N/A |
| N/A   N/A  N/A               N/A /  N/A | Not Supported        |     N/A          N/A |
|                                         |                      |                  N/A |
+-----------------------------------------+----------------------+----------------------+
                                                                                         
+---------------------------------------------------------------------------------------+
| Processes:                                                                            |
|  GPU   GI   CI        PID   Type   Process name                            GPU Memory |
|        ID   ID                                                             Usage      |
|=======================================================================================|
|  No running processes found                                                           |
+---------------------------------------------------------------------------------------+

我们关注机器的driver版本和cuda的版本如下

  • driver version: 540.4.0
  • cuda version: 12.6

此时我们进入cuda的release-notes中查看匹配范围

https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html

image.png

可以看到，我的内核在540版本，它不适合安装cuda 13.x， 而我自己版本安装的是12.6，符合cuda的要求

本地测试

默认情况下，12.6版本的cuda的sample程序没有预置，需要自己从github上安装，如下

git clone https://github.com/NVIDIA/cuda-samples.git

默认通过sdkmanager安装的版本，nvjpeg没有替我们安装，这里手动安装一下，方便后续直接编译

apt install libnvjpeg-dev-12-6
apt install g++ freeglut3-dev build-essential libx11-dev libxmu-dev libxi-dev libglu1-mesa-dev libfreeimage-dev libglfw3-dev

编译cuda的测试用例，需要为其指明cuda的安装位置，否则出现如下错误

CMake Error at CMakeLists.txt:5 (project):
  No CMAKE_CUDA_COMPILER could be found.

  Tell CMake where to find the compiler by setting either the environment
  variable "CUDACXX" or the CMake cache entry CMAKE_CUDA_COMPILER to the full
  path to the compiler, or to the compiler name if it is in the PATH.

由如下宏定义实现

CMAKE_CUDA_COMPILER=/usr/local/cuda-12.6/bin/nvcc

所以构建命令如下

cmake -DCMAKE_CUDA_COMPILER=/usr/local/cuda-12.6/bin/nvcc ..

最后，因为cuda-sample是针对所有的nvidia显卡的测试用例，为了能够在我们orin nano 上正常使用，我们需要先查询一下其代号，如下

https://developer.nvidia.com/cuda-gpus

image.png

可以看到，我这里使用的jetson orin nano 属于8.7的Compute Capability

所以在代码里面，我们应该指定为87，如下

set(CMAKE_CUDA_ARCHITECTURES 50 52 60 61 70 72 75 80 86 87 89 90 100 101 120)

如果是其他不支持的平台，例如compute_100，我们可以删除之。

测试deviceQuery

deviceQuery的代码在Samples/1_Utilities/deviceQuery，完成上述的预备操作之后，进入代码，直接编译即可。

mkdir build && cmake .. 
make

Consolidate compiler generated dependencies of target deviceQuery
[100%] Built target deviceQuery

此时测试程序编译出来了，我们运行即可

# ./deviceQuery 
./deviceQuery Starting...

 CUDA Device Query (Runtime API) version (CUDART static linking)

Detected 1 CUDA Capable device(s)

Device 0: "Orin"
  CUDA Driver Version / Runtime Version          12.6 / 12.6
  CUDA Capability Major/Minor version number:    8.7
  Total amount of global memory:                 7620 MBytes (7990489088 bytes)
  (008) Multiprocessors, (128) CUDA Cores/MP:    1024 CUDA Cores
  GPU Max Clock rate:                            1020 MHz (1.02 GHz)
  Memory Clock rate:                             1020 Mhz
  Memory Bus Width:                              128-bit
  L2 Cache Size:                                 2097152 bytes
  Maximum Texture Dimension Size (x,y,z)         1D=(131072), 2D=(131072, 65536), 3D=(16384, 16384, 16384)
  Maximum Layered 1D Texture Size, (num) layers  1D=(32768), 2048 layers
  Maximum Layered 2D Texture Size, (num) layers  2D=(32768, 32768), 2048 layers
  Total amount of constant memory:               65536 bytes
  Total amount of shared memory per block:       49152 bytes
  Total shared memory per multiprocessor:        167936 bytes
  Total number of registers available per block: 65536
  Warp size:                                     32
  Maximum number of threads per multiprocessor:  1536
  Maximum number of threads per block:           1024
  Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
  Max dimension size of a grid size    (x,y,z): (2147483647, 65535, 65535)
  Maximum memory pitch:                          2147483647 bytes
  Texture alignment:                             512 bytes
  Concurrent copy and kernel execution:          Yes with 2 copy engine(s)
  Run time limit on kernels:                     No
  Integrated GPU sharing Host Memory:            Yes
  Support host page-locked memory mapping:       Yes
  Alignment requirement for Surfaces:            Yes
  Device has ECC support:                        Disabled
  Device supports Unified Addressing (UVA):      Yes
  Device supports Managed Memory:                Yes
  Device supports Compute Preemption:            Yes
  Supports Cooperative Kernel Launch:            Yes
  Supports MultiDevice Co-op Kernel Launch:      Yes
  Device PCI Domain ID / Bus ID / location ID:   0 / 0 / 0
  Compute Mode:
     < Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 12.6, CUDA Runtime Version = 12.6, NumDevs = 1
Result = PASS

可以看到这里直接查询到板卡的所有基本信息。

测试conv_sample

conv_sample 是 jetpack 预装的示例包libcudnn9-samples，我们构建此用来来测试cuda，如下

cp -r /usr/src/cudnn_samples_v9/conv_sample/ /tmp/
cd /tmp/conv_sample && make

编译完成之后，直接运行测试，测试结果如下

# ./conv_sample 
Executing: conv_sample
Using format CUDNN_TENSOR_NCHW (for INT8x4 and INT8x32 tests use CUDNN_TENSOR_NCHW_VECT_C)
Testing single precision
====USER DIMENSIONS====
input dims are 1, 32, 4, 4
filter dims are 32, 32, 1, 1
output dims are 1, 32, 4, 4
====PADDING DIMENSIONS====
padded input dims are 1, 32, 4, 4
padded filter dims are 32, 32, 1, 1
padded output dims are 1, 32, 4, 4
Testing conv
^^^^ CUDA : elapsed = 0.00111403 sec,  
Test PASSED
Testing half precision (math in single precision)
====USER DIMENSIONS====
input dims are 1, 32, 4, 4
filter dims are 32, 32, 1, 1
output dims are 1, 32, 4, 4
====PADDING DIMENSIONS====
padded input dims are 1, 32, 4, 4
padded filter dims are 32, 32, 1, 1
padded output dims are 1, 32, 4, 4
Testing conv
^^^^ CUDA : elapsed = 0.0234335 sec,  
Test PASSED

可以看到测试cudann的转换成功。

测试tensorrt

jetpack默认预装了libnvinfer-samples，这是tensorrt的测试程序。下面基于此进行测试验证 关于tensorrt的测试说明，可以参考如下链接

https://docs.nvidia.com/deeplearning/tensorrt/latest/getting-started/quick-start-guide.html

这里以mnist的测试模型为例，展示通过onnx转换成trt的engine。借助libnvinfer-samples默认的测试程序，路径如下

/usr/src/tensorrt/samples/sampleOnnxMNIST

进入上述目录之后，直接make即可 编译完成之后，/usr/src/tensorrt/bin下会存在程序sample_onnx_mnist，直接运行即可验证

# ./sample_onnx_mnist
&&&& RUNNING TensorRT.sample_onnx_mnist [TensorRT v100300] # ./sample_onnx_mnist
[08/09/2025-18:53:25] [I] Building and running a GPU inference engine for Onnx MNIST
[08/09/2025-18:53:25] [I] [TRT] [MemUsageChange] Init CUDA: CPU +13, GPU +0, now: CPU 31, GPU 2505 (MiB)
[08/09/2025-18:53:28] [I] [TRT] [MemUsageChange] Init builder kernel library: CPU +927, GPU +583, now: CPU 1001, GPU 3133 (MiB)
[08/09/2025-18:53:28] [I] [TRT] ----------------------------------------------------------------
[08/09/2025-18:53:28] [I] [TRT] Input filename:   ../data/mnist/mnist.onnx
[08/09/2025-18:53:28] [I] [TRT] ONNX IR version:  0.0.3
[08/09/2025-18:53:28] [I] [TRT] Opset version:    8
[08/09/2025-18:53:28] [I] [TRT] Producer name:    CNTK
[08/09/2025-18:53:28] [I] [TRT] Producer version: 2.5.1
[08/09/2025-18:53:28] [I] [TRT] Domain:           ai.cntk
[08/09/2025-18:53:28] [I] [TRT] Model version:    1
[08/09/2025-18:53:28] [I] [TRT] Doc string:       
[08/09/2025-18:53:28] [I] [TRT] ----------------------------------------------------------------
[08/09/2025-18:53:28] [I] [TRT] Local timing cache in use. Profiling results in this builder pass will not be stored.
[08/09/2025-18:53:30] [I] [TRT] Detected 1 inputs and 1 output network tensors.
[08/09/2025-18:53:30] [I] [TRT] Total Host Persistent Memory: 18976
[08/09/2025-18:53:30] [I] [TRT] Total Device Persistent Memory: 0
[08/09/2025-18:53:30] [I] [TRT] Total Scratch Memory: 0
[08/09/2025-18:53:30] [I] [TRT] [BlockAssignment] Started assigning block shifts. This will take 4 steps to complete.
[08/09/2025-18:53:30] [I] [TRT] [BlockAssignment] Algorithm ShiftNTopDown took 0.031107ms to assign 2 blocks to 4 nodes requiring 31744 bytes.
[08/09/2025-18:53:30] [I] [TRT] Total Activation Memory: 31744
[08/09/2025-18:53:30] [I] [TRT] Total Weights Memory: 25704
[08/09/2025-18:53:30] [I] [TRT] Engine generation completed in 1.73945 seconds.
[08/09/2025-18:53:30] [I] [TRT] [MemUsageStats] Peak memory usage of TRT CPU/GPU memory allocators: CPU 0 MiB, GPU 8 MiB
[08/09/2025-18:53:30] [I] [TRT] [MemUsageStats] Peak memory usage during Engine building and serialization: CPU: 1437 MiB
[08/09/2025-18:53:30] [I] [TRT] Loaded engine size: 0 MiB
[08/09/2025-18:53:30] [I] [TRT] [MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +0, now: CPU 0, GPU 0 (MiB)
[08/09/2025-18:53:30] [I] Input:
[08/09/2025-18:53:30] [I] @@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@#-:.-=@@@@@@@@@@@@@@
@@@@@%=     . *@@@@@@@@@@@@@
@@@@%  .:+%%% *@@@@@@@@@@@@@
@@@@+=#@@@@@# @@@@@@@@@@@@@@
@@@@@@@@@@@%  @@@@@@@@@@@@@@
@@@@@@@@@@@: *@@@@@@@@@@@@@@
@@@@@@@@@@- .@@@@@@@@@@@@@@@
@@@@@@@@@:  #@@@@@@@@@@@@@@@
@@@@@@@@:   +*%#@@@@@@@@@@@@
@@@@@@@%         :+*@@@@@@@@
@@@@@@@@#*+--.::     +@@@@@@
@@@@@@@@@@@@@@@@#=:.  +@@@@@
@@@@@@@@@@@@@@@@@@@@  .@@@@@
@@@@@@@@@@@@@@@@@@@@#. #@@@@
@@@@@@@@@@@@@@@@@@@@#  @@@@@
@@@@@@@@@%@@@@@@@@@@- +@@@@@
@@@@@@@@#-@@@@@@@@*. =@@@@@@
@@@@@@@@ .+%%%%+=.  =@@@@@@@
@@@@@@@@           =@@@@@@@@
@@@@@@@@*=:   :--*@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@@@@@@@@@@@@@@@@@@@@@@@@@@@

[08/09/2025-18:53:30] [I] Output:
[08/09/2025-18:53:30] [I]  Prob 0  0.0000 Class 0: 
[08/09/2025-18:53:30] [I]  Prob 1  0.0000 Class 1: 
[08/09/2025-18:53:30] [I]  Prob 2  0.0000 Class 2: 
[08/09/2025-18:53:30] [I]  Prob 3  1.0000 Class 3: **********
[08/09/2025-18:53:30] [I]  Prob 4  0.0000 Class 4: 
[08/09/2025-18:53:30] [I]  Prob 5  0.0000 Class 5: 
[08/09/2025-18:53:30] [I]  Prob 6  0.0000 Class 6: 
[08/09/2025-18:53:30] [I]  Prob 7  0.0000 Class 7: 
[08/09/2025-18:53:30] [I]  Prob 8  0.0000 Class 8: 
[08/09/2025-18:53:30] [I]  Prob 9  0.0000 Class 9: 
[08/09/2025-18:53:30] [I] 
&&&& PASSED TensorRT.sample_onnx_mnist [TensorRT v100300] # ./sample_onnx_mnist

可以看到tensorrt的转换测试也是没有问题的

测试vpi

nvidia提供了vpi接口，默认也是通过jetpack的vpi3-samples提供测试示例，同样的也进行一下测试验证

cp -r /opt/nvidia/vpi3/samples/01-convolve_2d /tmp
cd /tmp/samples/01-convolve_2d && mkdir build && cmake .. && make

此时我们运行测试程序即可

./vpi_sample_01_convolve_2d cuda /opt/nvidia/vpi3/samples/assets/kodim08.png 
./vpi_sample_01_convolve_2d cpu /opt/nvidia/vpi3/samples/assets/kodim08.png

运行之后，本地存在两张图片，如下

edges_cpu.png  edges_cuda.png

两个图片一致即可。

使用容器

jetson orin 支持容器镜像运行，下面通过容器验证。

在使用之前，需要手动安装容器镜像，如下确定

apt install nvidia-container nvidia-container-runtime nvidia-container-toolkit nvidia-container-toolkit-base libnvidia-container1 nvidia-docker2

上面安装完成之后，需要确定docker的配置，如下

# cat /etc/docker/daemon.json
{
    "default-runtime": "nvidia",
    "runtimes": {
        "nvidia": {
            "args": [],
            "path": "nvidia-container-runtime"
        }
    }
}

如果设置不对，则需要手动写进去。

安装镜像

在安装容器镜像之前，需要先确定容器的版本和机器的版本尽可能一致

先查看机器的安装版本，通过jtop如下

image.png

可以看到我的板卡L4T的版本是36.4.4。所以我们找最新的容器版本，如下

https://catalog.ngc.nvidia.com/orgs/nvidia/containers/l4t-jetpack/tags

image.png

这里最新的是r36.4.0，点击右侧的复制，可以获取容器的镜像位置，然后如下进行运行容器

docker run -it --rm --net=host --runtime nvidia -e DISPLAY=$DISPLAY -v /tmp/.X11-unix/:/tmp/.X11-unix -v /root/github:/root/github -v /tmp/samples/01-convolve_2d/:/root/01-convolve_2d -v /usr/src/cudnn_samples_v9/conv_sample:/root/conv_sample nvcr.io/nvidia/l4t-jetpack:r36.4.0

这里值得注意的是，需要将主机运行的测试程序，通过-v选项，挂载到docker内运行。

测试deviceQuery

直接容器里面运行即可

/root/github/cuda-samples/build/Samples/1_Utilities/deviceQuery/deviceQuery

测试conv_sample

直接容器里面运行即可

/root/conv_sample/conv_sample

测试tensorrt

这个默认在jetpack中，直接编译运行

cd /usr/src/tensorrt/samples/sampleOnnxMNIST && make -j8
/usr/src/tensorrt/bin/sample_onnx_mnist

测试vpi

为了避免编译，直接运行如下

/root/01-convolve_2d/build/vpi_sample_01_convolve_2d cuda /opt/nvidia/vpi3/samples/assets/kodim08.png
/root/01-convolve_2d/build/vpi_sample_01_convolve_2d cpu /opt/nvidia/vpi3/samples/assets/kodim08.png

参考文献

https://catalog.ngc.nvidia.com/orgs/nvidia/containers/l4t-jetpack
http://forecr.io/blogs/programming/nvidia-container-runtime-1-installation
https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html
https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html

总结

至此，完成了一些基本的cuda的验证。